1 Authors and Affiliations

Sheena E. Martenies\(^a\), Sherry WeMott\(^a\), Grace Kuiper\(^a\), Kacy Lorber\(^b\), Cody Dawson\(^b\), Kevin Andresen\(^b\), William B. Allshouse\(^b\), Anne P. Starling\(^{c,d}\), John L. Adgate\(^b\), Dana Dabelea\(^{c,d,e}\), and Sheryl Magzamen\(^{a,c}\)

\(^a\)Environmental and Radiological Health Sciences, Colorado State University, Fort Collins, CO USA

\(^b\)Department of Environmental and Occupational Health, Colorado School of Public Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA

\(^c\)Department of Epidemiology, Colorado School of Public Health, University of Colorado Anschutz Medical Campus, Aurora, CO, USA

\(^d\)Lifecourse Epidemiology of Adiposity and Diabetes (LEAD Center), University of Colorado Anschutz Medical Campus, Aurora CO, USA

\(^e\)Department of Pediatrics, School of Medicine, University of Colorado Anschutz Medical Campus, Aurora, CO, USA

NOTE: The results of this analysis have not undergone peer-review and should be considered preliminary. Please do not cite or distribute.

The code for this project can be found on the author’s GitHub page here.

2 Abstract

Background

Denver, Colorado, a large metropolitan area at the base of the Rocky Mountains (elevation: 1609 m), suffers from poor air quality. Traffic is thought to be the predominant pollution source, but agriculture and industry also contribute. Our goal was to develop a land use regression (LUR) model to predict black carbon (PMBC) concentrations across the study area to facilitate retrospective exposure assessments for an ongoing birth cohort study.

Methods

Four seven-week campaigns were conducted to capture filter-based integrated black carbon (PMBC) measurements from May 2018 to March 2019. Weekly integrated PM2.5 samples were collected using low cost monitors and PMBC was measured using transmissometry at 880 nm. Weekly time-weighted average concentrations were calculated for each filter and concentrations were calibrated using a predictive model developed from a subset of filters co-located with a Federal Reference Method aethalometer. Filter concentrations were averaged to monthly means at each sampling location. Land-use characteristics were identified and summarized for buffer distances ranging 50-2500 m around each sampling site. We also included BC data from the local monitoring network and variables indicating the presence of wildfire smoke. PMBC concentrations were log-transformed prior to model fitting. Covariates were selected using a two-step process. First, a generalized linear model was fit using the least absolute shrinkage and selection operator (LASSO). Second, a linear model was fit using a stepwise AIC process and the LASSO-selected variables. The final model was validated using leave-one-out cross validation and used to predict monthly concentrations for a 250 m grid across the study area.

Results

We collected n = 917 filters from > 50 locations across the study area. The mean (SD) monthly PMBC concentration across all sampling locations was 1.32 (0.2) \(\mu\)g/\(m^2\). As expected, PMBC concentrations were highest near major roads and highest during the winter and summer months, consistent with traffic and biomass burning (residential wood stoves and wildfires) being the predominant sources of PMBC in the area. The final linear regression model for PMBC selected using the LASSO and Stepwise-AIC methods had 11 predictors. When ranked based on the t-statistic for each predictor, the five most important predictors were annual average daily traffic (AADT) in a 50 m buffer, month, temperature, average impervious surface in a 1000 m buffer, and population density in a 2500 m buffer. The final linear regression model had a adjusted R\(^2\) of 0.75 and a LOOCV R\(^2\) of 0.74.

Conclusions

Our LUR reasonably predicted PMBC concentrations in 2018 and 2019. Future work will focus on hindcasting the LUR-based predictions to retrospectively assign air pollutant exposures for an ongoing birth cohort study based in Denver, CO.

3 Introduction

Denver, Colorado, a major metropolitan area at the foot of the Rocky Mountains has a combination of topography, meteorology, and sources that create a unique pollution mix (Vedal et al., 2009). Traffic is a predominant source of air pollution, and as a result, the area is currently in non-attainment of the 2008 ozone National Ambient Air Quality Standard (NAAQS) (US EPA, 2018). Despite a recent history of poor air quality in the region, Denver, CO remains understudied with respect to the health effects of ambient air pollutants. The lack of air pollution epidemiology studies in Denver is partially due to a lack of robust models available to assess air pollutant exposures.

Land use regression (LUR) is a state-of-the-art method for modeling intra-urban air pollutant exposures, particularly for traffic related air pollution (TRAP). Over the last decade, black carbon (PMBC) has emerged as an important indicator of traffic-related air pollution (TRAP) exposures, complementing other indicators such as nitrogen dioxide and fine particulate matter with an aerodynamic diameter less than 2.5 \(\mu\)m (PM~2.5) (Hoek, 2017; Janssen et al., 2011). To address the gaps in available data for the Denver metropolitan area, we developed an LUR model focusing on PMBC as the pollutant of interest.

4 Methods

4.1 Study Area

The study area consists of the Denver Metropolitan area (Figure 1). In order to facilitate the development of a land use regression model for this area, the boundary was designed to capture most of the local sources of air pollution in the Denver metro area, including the busiest interstate highways (I-70 and I-25) and several large industrial facilities.

Figure 1. Sampling locations

Figure 1. Sampling locations

4.2 Sampling of PM2.5 in Denver, CO

Our filter-based sampling campaign was designed to capture temporal variability in traffic-related air pollutants in Denver, CO. We conducted four 7-week sampling campaigns during which low-cost monitors (Ultrasonic Personal Air Samplers [UPAS], AST, Fort Collins, CO) collected PM2.5 samples from n = 54 locations across the study area. These monitors were previously validated against a federal Reference Monitor (Volckens et al., 2017). The UPAS monitors were run at a 75% duty cycle and a flow rate of 1 L/min for at least 5 days. Our campaigns collected samples during each season and captured wildfire events in the summer and fall of 2018. Sampling locations were chosen to be representative of the land use characteristics for the study area (e.g., based on proximity to roads and traffic counts and percent impervious surface in a buffer around the site) to provide spatial variability in concentrations.

Samples were collected on Teflon filters and analyzed for PMBC (SootScan; Magee Scientific, Berkeley, CA) and metals (x-ray fluoroscopy). Based on the empirical elemental carbon relations between mass and absorbance for Teflon filters, we used \(\sigma\) values of 4.2 \(\mu\)g/\(cm^2\) to calculate the mass of PMBC based on absorbance at 880 nm. For both the SootScan and metals analysis, we used a standard filter area of 7.065 \(cm^2\).

4.3 Time-weighted average PMBC

We calculated time-weighted average (TWA) PMBC concentrations using the mass of PMBC measured on each filter, the sampled volume calculated from the flow rate (1 L/min) and the recorded run time for each sampler.

One UPAS monitor was co-located with a regulatory aethalometer maintained by the Colorado Department of Health and Environment (CDPHE). We compared our filter-based PMBC concentrations to the average PMBC concentrations recorded at the aethalometer for each filter collection period. Filter-based concentrations were calibrated using the time-weighted average PMBC and temperature recorded at the CDPHE monitor. After filter-based PMBC concentrations were calibrated, we averaged filters at each location to monthly means.

4.4 GIS-based predictors of TRAP exposure

We selected our GIS-based predictors of traffic-related air pollution on previous studies (e.g. Hankey and Marshall, 2015) and location specific indicators (e.g., wildfire smoke). The candidate predictors are summarized in Table 1. Predictors could either be spatial (e.g., land use characterization) or spatiotemporal (e.g., concentration of PM2.5 recorded at the closest FRM monitor). For non-distance related predictors (e.g., percent impervious surface), we started with six different buffers: 50 m, 100 m, 250 m, 500 m, 1000 m, and 2500 m. Distance and area-based GIS covariates were summarized for each sampling location in R.

Table 1. Summary of candidate LUR

Variable Type Metric Unit Buffer Distances (m) Source Temporal Resolution Spatial Resolution
Elevation Spatial Mean m 50, 100, 250, 500, 1000, 2500 US Geological Survey 100 m
Percent tree cover Spatial Mean % 50, 100, 250, 500, 1000, 2500 National Land Cover Database 30 m
Percent impervious surface Spatial Mean % 50, 100, 250, 500, 1000, 2500 National Land Cover Database 30 m
Land use classification Spatial Mode 50, 100, 250, 500, 1000, 2500 National Land Cover Database 30 m
Population density Spatial Mean persons / sq. km 50, 100, 250, 500, 1000, 2500 Socioeconomic Data and Applications Center 30 sec
Population count Spatial Mean persons 50, 100, 250, 500, 1000, 2500 Socioeconomic Data and Applications Center 30 sec
Distance to the closest airport Spatial m US Census Bureau Point
Distance to the closest CAFO Spatial m Colorado Department of Public Health and Environment Point
Distance to the closest compost facility Spatial m Colorado Department of Public Health and Environment Point
Distance to the closest highway Spatial m Colorado Department of Transportation Point
Distance to the closest landfill Spatial m Colorado Department of Public Health and Environment Point
Distance to the closest major road Spatial m Colorado Department of Transportation Point
Distance to the closest military installation Spatial m US Census Bureau Point
Distance to the closest mine Spatial m Colorado Department of Public Health and Environment Point
Distance to the closest National Priority List site Spatial m Colorado Department of Public Health and Environment Point
Distance to the closest oil/gas well Spatial m Colorado Oil and Gas Conservation Comission Point
Distance to the closest park or open space Spatial m US Census Bureau Point
Distance to the closest rail line Spatial m US Census Bureau Point
Distance to the closest wastewater treatment plant Spatial m Colorado Department of Public Health and Environment Point
Length of highways Spatial Sum m 50, 100, 250, 500, 1000, 2500 Colorado Department of Transportation Point
Length of major roads Spatial Sum m 50, 100, 250, 500, 1000, 2500 Colorado Department of Transportation Point
Annual average daily traffic Spatial Mean vehicles / day 50, 100, 250, 500, 1000, 2500 National Highway Performance Monitoring System Roadway Links
PM2.5 at closest monitor Spatiotemporal Mean µg/m³ US Environmental Protection Agency 24 hour Point
PM2.5 at closest 3 monitor Spatiotemporal Mean µg/m³ US Environmental Protection Agency 24 hour Point
PMBC at closest monitor Spatiotemporal Mean µg/m³ US Environmental Protection Agency 24 hour Point
Temperature at closest monitor Spatiotemporal Mean °F US Environmental Protection Agency 24 hour Point
Temperature at 3 closest monitors Spatiotemporal Mean °F US Environmental Protection Agency 24 hour Point
Smoke day at the closest monitor Spatiotemporal US EPA; NOAA Hazard Mapping System 24 hour Point
Smoke day at the 3 closest monitors Spatiotemporal US EPA; NOAA Hazard Mapping System 24 hour Point

4.5 Spatiotemporal predictors of PMBC

In addition to non-time-varying predictors such as land use characteristics, we also included several spatiotemporal predictors of PMBC. These included the monthly mean PM2.5, PMBC, and temperature recorded at the closest EPA monitoring site and an indicator of wildfire smoke. The presence of wildfire smoke was identified for each day of our sampling campaigns using data from the EPA monitoring network and the Hazard Mapping System from the National Oceanic and Atmospheric Administration (NOAA US EPA, 2019) using methods similar to those reported by Brey and Fischer (2016). We considered a monitor in the study area to be “wildfire smoke impacted” on a given day if the daily mean PM2.5 concentration measured at the monitor was more than 1 standard deviation above the 10 year average monthly concentration and if a wildfire smoke plume (as identified by NOAA analysts) was located within 50 km of the monitor. For the LUR, a sampling location was considered wildfire smoke-impacted for the month of interest if the closest monitor experienced any wildfire smoke days that month.

4.6 LUR Model Building

4.6.1 Predictor Pre-screening

Out combination of potential BC predictors and buffer distances resulting in 78 candidate covariates for our LUR model. However, previous work has found that LUR models that rely on fewer than 100 sites may be biased when using more than 20 predictors in the model (Basagana et al., 2013; Harrell et al., 1996). Therefore, we performed a pre-screening step to reduce the number of predictors used in model building. For each predictor for which we had multiple buffers, we fit single-predictor linear regression models to identify which buffer distance was most strongly associated with filter PMBC concentrations. Variables at different buffer distances were ranked by their R\(^2\) values, and we chose the buffer distance with the highest R\(^2\) for the model building process (discussed next). This pre-screening process reduced the number of potential covariates from 78 to 30.

4.6.2 Final model fitting

Out final model fitting was done using methods similar to those reported by Mercer et al. (2011). We used the least absolute shrinkage and selection operator (LASSO) to further reduce the covariates including in our LUR model. We implemented LASSO using the caret package in R. To assess the robustness of our LASSO-based model, We compared the models selected by LASSO to those selected by the RIDGE and elastic net algorithms and found that each approach resulted in similar median R\(^2\) and root mean square error (RMSE) values. To ensure the most parsimonious model was used, we compared the linear model with the LASSO-selected predictors to an alternative model fit using stepwise-AIC methods starting with the LASSO-selected predictors. Final model selection (Stepwise-AIC vs. LASSO) was made using AIC and adjusted-R\(^2\) as criteria. After model fitting, we confirmed that the assumptions for linear regression were met.

To evaluate the performance of our LUR, we used leave-one-out cross validation (LOOCV). We also calculated the root mean squared error (RMSE) for model predictions.

5 Results

5.1 Filter-based PMBC

We collected n = 917 filters from across more than 50 sampling locations across the Denver metro area between May 2018 and March 2019. The filter-based PMBC was calibrated using the following equation: \(Y_{calibrated} = 2.46 + 0.20 Y_{UPAS} - 0.05 Temp + 0.0004 {Temp}^2\). This calibration equation had an adjusted R\(^2\) value of 0.23. The mean (SD) monthly PMBC concentration across all filters was 1.32 (0.19) \(\mu\)g/m\(^3\) (Table 2). As expected, PMBC concentrations were highest for sampling locations closest to the highways.

Table 2. Summary statistics for monthly PMBC, stratified by quantiles of distance to highways

All filters (N = 404) > 1530 m (N = 99) 760 to 1530 m (N = 101) 281 to 759 m (N = 102) <= 280 m (N = 102)
BC (µg/m³)               
   mean (SD) 1.32 ± 0.19 1.26 ± 0.15 1.31 ± 0.17 1.32 ± 0.17 1.38 ± 0.23
   min 0.93 0.93 1.00 1.04 1.01
   max 2.1 1.84 1.77 1.95 2.10

5.2 Final LUR model for PMBC

Overall, the final model with 11 predictors performed similarly to the model with the LASSO-selected covariates; we selected this model to reduce the overall number of covariates included (Table 3). The final linear model had an adjusted R\(^2\) of 0.75 and a LOOCV R\(^2\) of 0.74. The RMSE for the final LUR model was 0.09 \(\mu\)g/m\(^3\).

Table 3. Comparison of candidate regression models for PMBC

Model No. Predictors AIC Adjusted R² LOOCV R² RMSE
AADT Only 1 -311.5706 0.263 0.258 0.162
All Predictors 30 -721.8132 0.761 0.722 0.087
LASSO Predictors 19 -718.1180 0.753 0.735 0.090
Stepwise-AIC Predictors 11 -733.7833 0.756 0.738 0.091

The coefficients for the final LUR model are summarized in Table 3. When ranked by their t-statistics, the five most important predictors of PMBC were annual average daily traffic (AADT) in a 50 m buffer, month, temperature, average impervious surface in a 1000 m buffer, and population density in a 2500 m buffer.

Table 4. Final linear regression model for PMBC
  bc_ug_m3
Predictors Estimates CI p
Intercept 2.47 2.02 – 2.92 <0.001
Temperature (F) -0.01 -0.01 – -0.01 <0.001
Elevation (100 m buffer) -0.00 -0.00 – -0.00 <0.001
Impervious surface (%, 1000 m buffer) 0.00 0.00 – 0.00 <0.001
Population density (n/sq km; 2500 m buffer) -0.00 -0.00 – -0.00 <0.001
Distance to highway (m) -0.00 -0.00 – 0.00 0.122
Distance to NLP site (m) -0.00 -0.00 – -0.00 <0.001
Distance to oil/gas well (m) 0.00 -0.00 – 0.00 0.096
AADT (50 m buffer) 0.00 0.00 – 0.00 <0.001
Smoke day 0.04 0.00 – 0.08 0.037
February 0.08 0.03 – 0.12 <0.001
March 0.13 0.09 – 0.18 <0.001
May 0.17 0.06 – 0.27 0.003
June 0.28 0.16 – 0.40 <0.001
July 0.35 0.23 – 0.47 <0.001
August 0.46 0.34 – 0.57 <0.001
October 0.11 0.06 – 0.17 <0.001
November 0.15 0.11 – 0.20 <0.001
BC at nearest monitor (ug/m3) -0.01 -0.05 – 0.04 0.801
Observations 396
R2 / R2 adjusted 0.767 / 0.755

    Plotting monthly PMBC concentrations for the filters (Figure 2) shows reasonable agreement between the measured and predicted values.

Figure 2. Comparison of meausured and predicted monthly mean PM~BC~

Figure 2. Comparison of meausured and predicted monthly mean PMBC

5.3 Predicted PMBC for the study area

Predicted PMBC concentrations by month are summarized in Table 5. Average PMBC concentrations tended to be higher in the winter and summer months, consistent with the increased use of wood-burning stoves in the winter for household heating and increased wildfire smoke contributions in the summer (July and August).

Table 5. Summary statistics for predicted PMBC by month (Jan, 2018 to June, 2019)

Year Month Mean (SD) Min Median Max
2018 1 1.23 (0.08) 0.86 1.22 1.75
2018 2 1.36 (0.08) 0.99 1.35 1.88
2018 3 1.29 (0.08) 0.93 1.28 1.80
2018 4 1.1 (0.08) 0.72 1.10 1.61
2018 5 1.12 (0.08) 0.75 1.12 1.64
2018 6 1.15 (0.09) 0.76 1.14 1.68
2018 7 1.21 (0.08) 0.83 1.20 1.73
2018 8 1.34 (0.08) 0.98 1.34 1.86
2018 9 0.9 (0.08) 0.54 0.89 1.41
2018 10 1.21 (0.08) 0.84 1.21 1.73
2018 11 1.36 (0.09) 0.99 1.35 1.87
2018 12 1.18 (0.1) 0.89 1.20 1.42
2019 1 1.26 (0.09) 0.88 1.25 1.77
2019 2 1.38 (0.09) 0.99 1.38 1.91
2019 3 1.36 (0.08) 0.99 1.35 1.88
2019 5 1.26 (0.09) 0.87 1.26 1.79
2019 6 1.21 (0.08) 0.92 1.21 1.71

PMBC displayed the expected spatial patterns across the metropolitan area. Figure 3 shows the PMBC predicted for the 250 m grid for the months of January (A), April (B), July (C), and October (D), 2018.. For all four months shown, concentrations were highest near the major roadways and lowest in areas with open space. Temporal variability in PMBC was smaller than spatial variability (Table 5), consistent with the predominant influence of traffic on PMBC concentrations.

Figure 3. Predicted PM~BC~ concentrations for four months in 2018

Figure 3. Predicted PMBC concentrations for four months in 2018

6 Discussion

Our monthly LUR model for PMBC in the Denver metropolitan area performed reasonably well. Our adjusted R\(^2\) and LOOCV R\(^2\) of 0.75 and 0.74 were consistent with or outperformed those reported in the literature (Dons et al., 2013; Hankey and Marshall, 2015; Kerckhoffs et al., 2016; Minet et al., 2018; Montagne et al., 2015; Saraswat et al., 2013; Tripathy et al., 2019). Our model included some key location-specific predictors, including the presence of wildfire smoke and distance to oil and gas wells, emphasizing the need to consider local sources when developing land use regression models.

6.1 Limitations

There are some important limitations to note for when interpreting the results of this analysis. First, we were not able to sample each week of the year, and thus some temporal data are missing. Second, we were limited to residential and public sites where the security of our monitors could be assured. Although we selected locations to be representative of the land use characteristics of our study area, we may have missed some key sources and sinks.

6.2 Next steps

This LUR model is intended to describe current and retrospective exposures for an ongoing birth cohort study in Denver, CO (Healthy Start; PI: Dabelea; 5UH3OD023248). To facilitate these exposure assessments, we will be completing the following next steps:

  • Fit additional LUR models for other pollutants of interest, including reactive oxidative species and metals associated with brake and tire wear

  • Hindcast the LUR predictions using the static geographic variable and temporally-resolved data from the local monitoring network

  • Evaluate the performance of hindcasted LUR predictions. Wang et al. (2013) found that LUR models underpredict retrospective exposures when concentrations are decreasing over time. However, concentrations of traffic-related air pollutants (e.g., PM2.5 and NO2) measured at each monitor in the Denver metro area have remained roughly level over the last decade (US EPA, 2019) suggesting spatial patterns of exposure are similar between the start of the cohort recruitment period in 2009 and now

7 References

Basagaña X, Aguilera I, Rivera M, Agis D, Foraster M, Marrugat J, et al. 2013. Measurement error in epidemiologic studies of air pollution based on land-use regression models. Am J Epidemiol 178:1342–1346; doi:10.1093/aje/kwt127.

Brey SJ, Fischer EV. 2016. Smoke in the City: How Often and Where Does Smoke Impact Summertime Ozone in the United States? Environ Sci Technol 50:1288–1294; doi:10.1021/acs.est.5b05218.

Dons E, Van Poppel M, Kochan B, Wets G, Int Panis L. 2013. Modeling temporal and spatial variability of traffic-related air pollution: Hourly land use regression models for black carbon. Atmospheric Environment 74:237–246; doi:10.1016/j.atmosenv.2013.03.050.

Hankey S, Marshall JD. 2015. Land Use Regression Models of On-Road Particulate Air Pollution (Particle Number, Black Carbon, PM2.5, Particle Size) Using Mobile Monitoring. Environ Sci Technol 49:9194–9202; doi:10.1021/acs.est.5b01209.

Harrell JF, Lee KL, Mark DB. 1996. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in medicine 15: 361–387.

Hoek G. 2017. Methods for Assessing Long-Term Exposures to Outdoor Air Pollutants. Curr Envir Health Rpt 4:450–462; doi:10.1007/s40572-017-0169-5.

Janssen NAH, Hoek G, Simic-Lawson M, Fischer P, van Bree L, ten Brink H, et al. 2011. Black carbon as an additional indicator of the adverse health effects of airborne particles compared with PM10 and PM2.5. Environ Health Perspect 119:1691–1699; doi:10.1289/ehp.1003369.

Mercer LD, Szpiro AA, Sheppard L, Lindström J, Adar SD, Allen RW, et al. 2011. Comparing universal kriging and land-use regression for predicting concentrations of gaseous oxides of nitrogen (NOx) for the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air). Atmospheric Environment 45:4412–4420; doi:10.1016/j.atmosenv.2011.05.043.

Minet L, Liu R, Valois M-F, Xu J, Weichenthal S, Hatzopoulou M. 2018. Development and Comparison of Air Pollution Exposure Surfaces Derived from On-Road Mobile Monitoring and Short-Term Stationary Sidewalk Measurements. Environ Sci Technol 52:3512–3519; doi:10.1021/acs.est.7b05059.

Montagne DR, Hoek G, Klompmaker JO, Wang M, Meliefste K, Brunekreef B. 2015. Land Use Regression Models for Ultrafine Particles and Black Carbon Based on Short-Term Monitoring Predict Past Spatial Variation. Environ Sci Technol 49:8712–8720; doi:10.1021/es505791g.

Saraswat A, Apte JS, Kandlikar M, Brauer M, Henderson SB, Marshall JD. 2013. Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon Particulate Matter in New Delhi, India. Environ Sci Technol 47:12903–12911; doi:10.1021/es401489h.

Tripathy S, Tunno BJ, Michanowicz DR, Kinnee E, Shmool JLC, Gillooly S, et al. 2019. Hybrid land use regression modeling for estimating spatio-temporal exposures to PM2.5, BC, and metal components across a metropolitan area of complex terrain and industrial sources. Sci Total Environ 673:54–63; doi:10.1016/j.scitotenv.2019.03.453.

US Environmental Protection Agency [US EPA]. 2018. 8-Hour Ozone (2008) Designated Area/State Information with Design Values. Green Book. Available: https://www3.epa.gov/airquality/greenbook/hbtcw.html [accessed 26 June 2018].

US Environmental Protection Agency [US EPA]. 2019. AirData Website File Download Page. Available: https://aqs.epa.gov/aqsweb/airdata/download_files.html [accessed 3 July 2019].

Vedal S, Hannigan MP, Dutton SJ, Miller SL, Milford JB, Rabinovitch N, et al. 2009. The Denver Aerosol Sources and Health (DASH) Study: Overview and Early Findings. Atmos Environ 43:1666–1673; doi:10.1016/j.atmosenv.2008.12.017.

Volckens J, Quinn C, Leith D, Mehaffy J, Henry CS, Miller-Lionberg D. 2017. Development and evaluation of an ultrasonic personal aerosol sampler. Indoor Air 27:409–416; doi:10.1111/ina.12318.

Wang R, Henderson SB, Sbihi H, Allen RW, Brauer M. 2013. Temporal stability of land use regression models for traffic-related air pollution. Atmospheric Environment 64:312–319; doi:10.1016/j.atmosenv.2012.09.056.